I missed the first AI wave because I assumed it was for people who understand terminals. It isn’t.
I am an advocate in Johannesburg, about 25 years in. We have a referral profession here: I work for my own account and take instructions from attorneys, not directly from clients. But the work will look familiar to any litigator anywhere. Research, reading a very large record, drafting pleadings and affidavits, opinions, written argument, correspondence, and the endless job of working out what actually matters in a file.
For most of my career I did that work the traditional way. Law reports, textbooks, the databases, SAFLII. Download the judgment, copy the quote, organise the documents, produce the thing in Word and Acrobat. None of it glamorous. Most of it manual. But it taught me the thing that turns out to matter most with AI: know where a proposition came from before you rely on it.
I missed the first wave
I did not jump on ChatGPT or Gemini or Claude when they appeared. I am not a coder. I do not naturally understand a terminal, a command line, or what is going on behind a programme. I am a consumer of ordinary tools: websites, Word, Acrobat.
That left me in an awkward middle. A chat window could answer a question or tidy a paragraph, but it was obviously not part of my real working life. The powerful stuff seemed to belong to people who could code. I assumed it was not for me.
That changed early in 2026 when I came across Zack Shapiro’s writing about using Claude for legal work. The idea was unexpectedly simple. Rather than ask an AI a question in isolation, give it a properly prepared working environment (a folder, some reference material) and clear instructions. That was the bridge I had been missing. It did not need me to become a programmer. It did need me to become organised and precise, which was less fun but more achievable.
The preparation is the work
The most useful thing I did first was not a clever prompt. It was digging out my own old work.
Over the years I had accumulated settled examples of every recurring document: heads, pleadings, affidavits, opinions, memoranda. I put good examples of each into separate folders and asked Claude to read them and write me a style guide for each type. Not just the language and structure, but the boring bits: headings, fonts, spacing, how I number paragraphs.
This is not “training” in any technical sense. It is giving the tool something to copy. It means that when it produces a first draft, the draft has a chance of sounding like a document prepared for the right court, rather than the generic AI prose we all now recognise at fifty paces. The point is not to hand over the judgment. The point is a first draft I can properly consider, change, settle and own.
A small note about files ending in .md. They are Markdown files: plain text with headings and lists. Claude reads them easily and so do you. My style guides, matter notes and indexes are all .md. It is not magic. It is a text file.
Set up before you touch a real matter
Before a client file goes anywhere near the thing, tell it who you are and how you work. I keep a short standing instruction: I am a litigator, accuracy beats fluency, important propositions get a source, uncertainty is said out loud not hidden. I cannot tell you this makes it reliable. It will repeat your instructions back to you very reassuringly and then get something wrong anyway. It helps. It is not a guarantee. You still instruct per task, and you still check.
Be careful about access too. One folder per matter, with only the material you are entitled to use in it. Do not point the thing at your whole computer and assume you understand every permission that follows.
And a more basic warning, from my own bank statement. When I first went looking for Claude I landed on a site that looked plausible and paid about US$80 before I understood I had bought some intermediary’s product, not Anthropic’s. Use the official link. Read what you are buying. Be suspicious of the sponsored result at the top of the search.
PDFs are where confidence goes wrong
The most dangerous documents in a litigation workflow are the ordinary-looking ones. A PDF can be perfectly readable to you and mostly unreadable to a machine: a scan rather than real text, a broken text layer, a watermark, a table, handwriting, a stamp, a page the wrong way up. A long bundle will produce a fluent answer even where the tool has only managed to read part of it. I have had several near misses where I assumed a PDF had been read and it hadn’t. The whole story, and what I do about it now, is here, and there is a plain explanation of why it happens here.
Do your own research
This is the caution I think matters most. Keep finding the law the way you trust. The official court site, the database you subscribe to, the reports. Download the current rules and directives and keep them in a folder you control, and give the tool that folder.
That does not make AI useless for research. It can find issues, summarise a judgment you have supplied, compare two texts, build a chronology, list the questions, and draft something that points you back to the source. What it should not be is an oracle. Ask it for a case that helps your client and it may give you something that looks perfect: a convincing name, a neat citation, a plausible quote, exactly the proposition you wanted. It may still be fiction. The early warning was Mata v Avianca in the United States, where lawyers filed non-existent authorities generated by ChatGPT and were sanctioned. The lesson is not that AI is wicked. It is that a plausible answer is not evidence, and every authority still gets checked at the source.
The first task that shows what is possible
Once the reference material is in place, the matter folder is set up, the documents have been checked, and the instructions are written, ask for a real piece of work. The best mental model is a diligent junior who has the file but needs complete instructions. Do not say “prepare a letter of demand for discovery”. You will get something generic. Say what matter it is, who you act for, what stage it is at, what correspondence exists, which rule applies and must be quoted, what is required, by when, from what date, and whether court holidays affect it. Ask it to flag uncertainty, quote the rule exactly, and use the matter reference and your document style.
The result can feel like magic the first time. It is not. It is better material, better context and better instructions. It may save time. It may give you a surprisingly good first draft. It may change what a solo practitioner can take on. It does not replace the person whose name is on the opinion, the letter, the affidavit or the heads.
That person is still you.
Further reading: Mata v Avianca and Ayinde are essential reading on false authorities and professional responsibility. See the Research page. Next: what Claude looks like, and how to hand it a document.